Skip to content

Latest commit

 

History

History
25 lines (13 loc) · 1.7 KB

README.md

File metadata and controls

25 lines (13 loc) · 1.7 KB

A Computational Framework for Slang Generation - Code and Data Repository

This is the github repository for the TACL paper "A Computation Framework for Slang Generation".

Urban Dictionary (UD) Dataset for Slang NLP

The dataset is a curated subset of the open source Urban Dictionary subset made available via Kaggle:

https://www.kaggle.com/therohk/urban-dictionary-words-dataset

Please see our paper for details regarding how the entries were selected and processed.

You can find the dataset under the /UD_Dataset directory.

The .csv files contain the data in raw text format (UD_Data_full.csv) and are split into training (UD_Data_train.csv) and testing (UD_Data_test.csv) partitions.

UD_Dataset.npy contains a python friendly format of the dataset and also includes sample trained contrastive embeddings based on both fastText and SBERT along with the baseline variants without contrastive training. See the accompanied IPython notebook (UD_Dataset.ipynb) for an usage example.

Training Contrastive Sense Embeddings

Code for training contrastive embeddings can be found in the /Code directory. Please refer to the Jupyter notebook found in the /Demo directory for a tutorial.

The provided demo showcases the complete training procedure to obtain contrastive slang sense embeddings from the UD dataset released above along with open sourced conventional definitions from WordNet. This combination achieves comparable results as the UD experiment performed in the paper that uses Oxford Dictionary (OD) entries as conventional definitions.